`Hermitian()` wrapper missing a GPU compatible iterator

[abussy]

The following reproducer code fails with LoadError: Scalar indexing is disallowed.:

using CUDA
using LinearAlgebra

A = CuArray(ones(Float32, 5, 5))

# This works
any(isnan, Symmetric(A))

# This does not
any(isnan, Hermitian(A))

It seems that some underlying function call tries explicit element access with the Hermitian() wrapper, but not with Symmetric(). I am not 100% sure whether that issue belongs here or in LinearAlgebra.

I get this issue with Julia version 1.10.8 (LTS) and 1.11.3 (current stable).

[cpmores]

Seems that the optimized function 'any' cannot resolve Hermitian{T,S<:AbstractMatrix{<:T}} <: AbstractMatrix{T}.

'any' in mapreduce.jl looks like:

Base.any(f::Function, A::AnyGPUArray) = mapreduce(f, |, A)

But Hermitian or Symmetric is a struct including UP or Down, which cannot go into this optimized function.

If you use the data instead of the whole struct, any will work properly.

using CUDA
using LinearAlgebra

A = CuArray(ones(Float32, 5, 5))

# This works
S = Symmetric(A)
println(any(isnan, S.data))

# This works
H = Hermitian(A)
println(any(isnan, H.data))


B = CuArray(fill(NaN, 5, 5))
# This works
SB = Symmetric(B)
println(any(isnan, SB.data))

# This works
HB = Hermitian(B)
println(any(isnan, HB.data))